- Fix a bug in eta-expansion of an application; it probably never bit
  us in practice.

- Change the behaviour of State#-lambdas: previously a State# lambda
  was considered to be one-shot for the purposes of inlining and
  eta-expansion.  Both can lead to unbounded extra work, but inlining
  is more likely to kill us, and eta-expansion is the case we mostly
  care about (getting the State# lambda to the outside of the function
  is often very important).  So now we only consider a State# lambda to be
  one-shot in eta expansion, not for inlining.